Meta-Dynamical State Space Models for Integrative Neural Data Analysis
Authors: Ayesha Vermani, Josue Nassar, Hyungju Jeon, Matthew Dowling, Il Memming Park
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We first validate the proposed method on synthetic data and then test our method on neural recordings from the primary motor and premotor cortex. We demonstrate the efficacy of our approach on few-shot reconstruction and forecasting of synthetic dynamical systems, and neural recordings from the motor cortex during different arm reaching tasks. Specifically, our approach outperformed other methods on forecasting observations with all methods having comparable reconstruction performance (Fig. 4B, C). We include further validation experiments when there is no model mismatch as well as the generalization of the trained model to new data in Appendix B. We additionally include results on these recordings from multi-session CEBRA (Schneider et al., 2023) in Appendix B. |
| Researcher Affiliation | Collaboration | 1 Champalimaud Centre for the Unknown, Champalimaud Foundation, Portugal 2 Ryvivy R, USA EMAIL |
| Pseudocode | No | The paper describes methods using mathematical equations and diagrams (e.g., Figure 14: Inference Overview) but does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states: "We used the official implementation of LFADS in Py Torch to obtain the results reported in the paper (Sedler & Pandarinath, 2023)." This refers to a third-party tool used, not an explicit release of the authors' own implementation code for the methodology described in this paper. |
| Open Datasets | Yes | We used single and multiunit neural population recordings from the motor and premotor cortex during two behavioral tasks the Centre-Out (CO) and Maze reaching tasks (Perich et al., 2018; Gallego et al., 2020; Churchland et al., 2012). |
| Dataset Splits | Yes | For all experiments, we split each of the M datasets into a training and test set and report reconstruction and forecasting metrics on the test set. For each value of ωi, we generated 128 latent trajectories for training, 64 for validation and 64 for testing, each of length T = 300... We report performance on 64 test trajectories from each dataset. We used 512 trials when available or 80 percent of the trials, each of length 36, for training all methods. For evaluating the decoding performance from forecasted observations, we used the first 13 time bins (around time till movement onset) to estimate the latent state and embedding, and subsequently use the trained dynamics model to forecast the next 20 time bins. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, memory, or specific computing cluster configurations used for running the experiments. |
| Software Dependencies | No | The paper mentions "Py Torch" in the context of using LFADS implementation, but no specific version number is provided for PyTorch or any other software component critical to replicating their method. |
| Experiment Setup | Yes | We used the Adam optimizer with weight decay and a Cosine annealing schedule on the learning rate for pretraining all approaches. We used the LAMB optimizer for pretraining all multi-session methods on the motor cortex recordings and used Adam with weight decay for the single-session models, with a Cosine annealing schedule on the learning rate in both cases. Synthetic Examples: lr: 0.005 weight decay: 0.001 batch size: 8 from each dataset. Motor Cortex Experiment: lr: 0.01 weight decay: 0.05 batch size: 64 trials from 20 datasets. Inference network: Ωi: MLP(dyi, 128, Dropout(0.6), 64) qα: [GRU(64), Linear(64, 2 de)] qβ: [bi GRU(128), Linear(128, 60)]. Generative model: fθ: MLP(30, 128, 128, 30) hϑ: MLP(de, 64, 64, (256 + 158) dr) pϕi: [Linear(30, dyi)] |